Abstract

Speech Emotion Recognition (SER) is considered a challenging task in the domain of Human-Computer Interaction (HCI) due to the complex nature of audio signals. To overcome this challenge, we devised a novel method to fine-tune Convolutional Neural Networks (CNNs) for accurate recognition of speech emotion. This research utilized the spectrogram representation of audio signals as input to train a modified Alex Net model capable of processing signals of varying lengths. The IEMOCAP dataset was utilized to identify multiple emotional states such as happy, sad, angry, and neutral from the speech. The audio signal was preprocessed to extract a 3D spectrogram that represents time, frequencies, and color amplitudes as key features. The output of the modified Alex Net model is a 256-dimensional vector. The model achieved adequate accuracy, highlighting the effectiveness of CNNs and 3D Mel-Spectrograms in achieving precise and efficient speech emotion recognition, thus paving the way for significant advancements in this domain.